How to tell if AI is really saving you money

An AI that reads an invoice in one minute instead of six saves five minutes on that task. Whether it saves your company money depends on what happens before and after that minute: who checks the result, how long the invoice waits until someone approves it and how many come back because of a mistake.

The short answer: to know whether AI is really saving you money, measure what each finished case costs, from start to finish, before and after you introduce it and under the same rules. The time saved on a single task almost never matches the money saved in the process.

Why the time saved on a task misleads

A process is a chain: a document comes in, someone reads it, someone records it, another person reviews it, a third approves it and it goes out. Speeding up one link speeds up the whole only if that link was the bottleneck. When the bottleneck is review or approval, the link AI speeds up is free sooner and the work piles up at the next one.

Studies that have measured this in a controlled way, in different settings, point the same way:

• Customer support, measured as the output of the job. Brynjolfsson, Li and Raymond followed 5,179 customer support agents who were given a generative AI assistant and measured issues resolved per hour, that is, what leaves the job and not the time of a single reply. The average improvement was 14%, and 34% among the least experienced workers (NBER, 2023). It is an example of a real saving, and it shows because what leaves the process was measured.
• Programming, where perception fails. In 2025 METR ran a randomised trial with 16 experienced developers and 246 real tasks on their own projects. With AI they took 19% longer, and still believed they had been 20% faster (METR, July 2025). The study used early-2025 tools, it cannot be applied to other kinds of work and METR has since published more recent data. What it leaves is something else: asking the team how much they think they save is not a measurement.
• Programming, where the bottleneck moves. Faros AI analysed telemetry from more than 10,000 developers in 1,255 teams. In teams with the most AI use, 21% more tasks were completed and 98% more changes were merged, but the review time of each change rose 91%, and they found no improvement in company-level indicators (Faros AI, July 2025). The data are correlational and published by a vendor in the sector.

None of the three is about invoices, case files or orders. The mechanism is the same: each person produces more on their own task, and the whole process delivers what its slowest step allows.

What to measure: five numbers per process

These five numbers are recorded for each case, before and after introducing AI:

The last one decides, and the other four explain why it goes up or down. For a prior estimate, the ROI calculator (in Spanish) gives a guide by sector, but as soon as you have your own data, measurement replaces any estimate.

How to check it in four weeks

1. Choose a process with a clear start and end and some volume. Supplier invoices, case files, orders or site reports work, provided there are dozens of cases a month. A job that happens twice a year does not.
2. Measure how it is today. For two weeks, record thirty or so real cases or more with the five numbers. Without this baseline there is nothing to compare with.
3. Put the AI in with a person who reviews every result and change nothing else at the same time, neither the team nor the order of the steps. Human oversight has levels, and the one you choose counts as work.
4. Measure for another two weeks under the same rules. A Tecnea proof of concept is done this way, with real documents from your company, in two to four weeks and from €850 (what each amount includes, in Spanish).
5. Decide with a threshold you set beforehand: how far the cost per case has to fall, or how much the weekly output has to rise, for it to be worth continuing. Setting it beforehand avoids deciding on what one hopes to see.

When a saving really counts

Freeing up hours is not saving money if nobody uses them for something else. A saving shows up in one of these three ways:

1. Fewer paid hours for the same work: less overtime, fewer temps, fewer contractors.
2. More work with the same team: more orders handled or more case files closed each month. It is growth more than saving, and it is measured in extra cases and in what each one contributes.
3. Fewer avoidable errors: returns, penalties, complaints. It is calculated as the cost of one error times the errors that no longer happen.

If none of the three appears in the numbers, the AI will have saved time on the task and money nowhere. That is a useful result: it avoids investing in a full development on a hypothesis that does not hold.

Common mistakes when measuring

• Measuring with a demo or with the easy cases. The sample has to include the exceptions, which is where the time goes.
• Not counting the review. If a person checks every result, those minutes are part of the cost per case.
• Forgetting the cost of the AI. Licences, AI usage, maintenance and the monthly fee all go into the cost per finished case.
• Trusting the team's perception. The METR study is the warning: what each person believes they save does not always match what was measured.
• Changing two things at once. If the team is reorganised and AI is introduced in the same month, nobody will know which to credit for the result.
• Measuring too early. In the first weeks the team is learning the tool, and it is worth letting that period pass before the measurement that counts.

What we measure ourselves

In the projects Tecnea puts into production, the results measured are those of the process. In a self-service portal for more than 50 companies, 95% fewer incidents and 70% fewer support queries. In Tecnea's own internal management, 80% less time on administrative tasks and seven applications replaced by one. These figures were measured in those projects (case studies, in Spanish) and hold for them: your process will have its own, which is why it is measured before deciding.

Frequently asked questions

How long does it take to know whether AI saves money?

With enough volume in the process, between four and six weeks: two or more for the baseline and another two for the test. With low volume it takes longer to gather enough cases, not less rigour.

Which process is best to measure first?

One that repeats often, has a clear start and end and works on documents: supplier invoices, orders, case files, site reports. There, the cost per case is easy to calculate. In what an AI agent already does in your sector there are examples by activity, and in the AI that comes with your software or custom development how to decide which of the two options to measure.

How do I turn saved time into euros?

By multiplying the hours that are no longer paid for (overtime, temps, contractors) by their cost, and adding the value of the extra cases and of the errors avoided. Freed hours that nobody reassigns do not count until they turn into one of those three things.

What if the numbers come out badly?

It is a valid result and a cheap one to obtain. A proof of concept exists for that: to know before investing in the full development. If it does not pay off, that has to be said too.

If you want to measure a specific process in your company before investing, tell us which one.

This article is informative. Tecnea develops custom applications with AI, so we have an interest in you measuring and, if it pays off, commissioning one: that is why we first explain how to check the saving before investing. Third-party figures state what they measure and in what context; the METR and Faros studies refer to software development and cannot be applied as they are to other processes.

Sources

1. Brynjolfsson, Li and Raymond, "Generative AI at Work", NBER Working Paper 31161 (2023): nber.org
2. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (10 July 2025): metr.org
3. Faros AI, "The AI Productivity Paradox Report 2025" (23 July 2025): faros.ai
4. Tecnea, case studies (in Spanish): figures measured in each project.